Open Weights: The Actual Trade

00

The four-part underwriting frame

A model can become cheaper without every downstream beneficiary retaining the savings—and without every token requiring a new accelerator.

Market structure

Two tiers, not one winner

Open weights win cheap volume. Closed frontier models retain high-stakes spend.

NVIDIA strategy

Models as complements

More deployable models expand demand for chips, networking, systems and software.

Capture test

Tokens ≠ new GPUs

Efficiency, old fleets, custom silicon and lower prices can absorb demand.

Policy

Open at home, targeted abroad

Domestic openness coexists with enforcement against alleged IP and compute diversion.

01

Volume is commoditizing faster than value

Vercel’s June production index is the constraint that keeps both narratives honest. Scope: Vercel Gateway traffic and market list prices, not the entire market or negotiated invoices.

Vercel AI Gateway / June 2026

The spend gap is the story.

Open-weight share of tokens29%

29%

Open-weight share of spend<4%

<4

Anthropic share of tokens32%

32%

Anthropic share of text spend61%

61%

Open models: ~1/10 average token price Closed frontier average price/token: +~12%

Adopt

Routing is the equilibrium.

Low-value, privacy-sensitive and latency-tolerant tasks move open. High-stakes coding and automation continue to pay a premium.

Victim test

Thin wrappers, not SaaS broadly.

Falling model COGS helps only if the vendor owns workflow, data or distribution and retains the savings.

Invalidation

Watch spend, not downloads.

The thesis changes when open weights take material premium-workload spend—not when model cards or download counts rise.

02

Adopt, modify, reject

The alternate write-up has the better incentive lens, but several hard-data claims fail live verification. This table is the reconciled record.

Claim	Decision	Final resolution
Signatories profit from model commoditization	Adopt	Read the list as an incentive map. Do not elevate it to proof of a formal anti-lab coalition.
No signatory sells a closed frontier model	Reject	Microsoft, Perplexity and Mistral complicate the claim. Better: no U.S. pure-play closed-frontier lab dependent primarily on scarcity signed.
Distillation is the load-bearing ask	Modify	It is the contested mechanism. The letter protects legitimate distillation but accepts targeted remedies for unlawful extraction.
NVDA −18% YTD / ~19× forward	Reject	Through 24 July: approximately $206.84, +~11% YTD, and ~23.8× Yahoo consensus forward P/E.
B200 rental −31% in three weeks	Reject	Silicon Data: $5.63/hour and +0.2% over seven days. Documented July provider change: up to −6% index impact.
OpenAI abandoned first-party Stargate	Reject	Some projects were trimmed, but a 3.2 GW Georgia campus was disclosed this week with greater OpenAI control of design.
DRAM +~90% in Q1 2026	Adopt	TrendForce: conventional DRAM +90–95% QoQ forecast; server DRAM around +90%. Q3 server forecast moderated to +13–18%.
Memory 40–50% of BOM / HBM 36–52 weeks	Exclude	Directional scarcity is credible. The exact current figures were not substantiated from sufficiently reliable sources.
Old fleets have a 3–4× cost advantage	Reject	A100 is ~3.4× cheaper per GPU-hour than B200, but hourly rate is not cost per useful token. Throughput, memory and interconnect differ.
Live policy points against open weights	Modify	Policy is bifurcated: support domestic open weights; target alleged malicious distillation, IP theft and compute diversion.
The letter is not policy	Adopt	No bill, sponsor, committee action or rulemaking. It is strategic lobbying and agenda-setting.

03

The NVIDIA capture equation

Open weights raise demand elasticity. The stock outcome still depends on who serves the tokens, on what hardware, at what utilization and margin.

Incremental NVIDIA value

DemandToken + deployment elasticity

×

CaptureNVIDIA share × new-hardware intensity

−

LeakageEfficiency + old fleets + custom silicon

−

EconomicsPrice pressure + supply-chain rent

Jevons for tokens is not automatically Jevons for next-generation NVIDIA revenue. Kimi K3 also shows why the high end can remain hardware-intensive: Moonshot recommends supernodes with at least 64 accelerators, but the promised full weights and independent self-host economics were not yet available on 25 July.

04

One government, two policy vectors

The apparent contradiction disappears when domestic open-weight policy is separated from China-specific enforcement and compute controls.

Domestic vector

Open, diverse, deployable

AI Action Plan supports open-source and open-weight AI.
EO 14409 uses a voluntary frontier-model framework rather than mandatory preclearance.
NSPM-11 encourages commercial and open-source models from diverse vendors.
NVIDIA’s letter asks for compute access, shared assets and no premature restrictions.

Coexist

Enforcement vector

Target IP and compute diversion

Bessent said sanctions are possible if Chinese-model IP theft is established.
Existing export-control and Entity List authorities can reach hardware access.
NSPM-11 explicitly addresses malicious distillation.
A Moonshot-specific action is not a ban on domestic open weights.

05

Where value is most likely to settle

Ranked by quality of exposure to the thesis—not by current valuation. Each still requires a separate company-level earnings and price discipline.

01 / Highest qualitySelective long

Workflow owners

Application and service companies that can route models, own proprietary data and retain COGS savings. Signatories are a research list, not a buy list.

NOW / PLTR / CRWD / BOX / IBM

02 / Direct contentCycle aware

HBM + server memory

Decode bandwidth, model weights and KV cache support structural content. Avoid calling it “underpriced” without earnings-cycle work.

MU / 000660 / 005930

03 / DeploymentStructural long

Networking + power

More inference increases rack-scale interconnect and energized-power demand. Underwrite live megawatts, not announced gigawatts.

Rack economics

04 / Core beneficiaryNot event trade

NVIDIA

Open models are strategic complements. Size on capex durability, share, gross margin and valuation—not on this letter.

NVDA

05 / Risk heavyAvoid factor shortcut

Neoclouds + wrappers

Compute lessors are financing structures; thin wrappers lose barriers as inputs cheapen. Neither is a clean open-weight beta.

CRWV / NBIS / thin apps

06

The priced-in book

Industry winners are not automatically stock buys. Reference tape: 24 July 2026 close; YTD from 31 December adjusted closes. Bands are entry disciplines, not price targets.

Portfolio conclusion

Buy delivered economics. Sell capitalized hope.

The best asymmetry is NVIDIA plus de-rated workflow control planes. The physical bottlenecks remain real, but MU and VRT already discount exceptional scarcity. The cleanest hedge is expensive catch-up and long-duration expectation—not “closed models” broadly.

Core longNVDA: +11% YTD versus +85% revenue growth

Funding sourceTrim MU + VRT after +223% / +79% YTD

Relative hedgeShort AMD + PLTR expectations, not the AI theme

Buy / core

NVDA

$206.84 · +11% YTD · 31.7× trailing / ~23.8× forward

What is priced: major AI growth plus deceleration. What is not: a wider open-model demand funnel with continued estimate upside. Add below $195–200; reassess above $245–255 if estimates flatten.

Buy / selective

IBM · NOW

−27% / −36% YTD · 19.0× / 61.7× trailing

IBM: value-priced neutral control plane. NOW: higher-quality workflow owner, still multiple-sensitive. Start NOW at one-third size; build IBM around $205–220.

Trim / wait

MU · VRT

+223% / +79% YTD · 20.9× / 72.9× trailing

The bottlenecks are real. The discovery trade is over. MU’s low P/E may be peak-cycle earnings; VRT already discounts sustained power scarcity. Keep tracking positions, not fresh full-size longs.

Sell / hedge

AMD · PLTR

+144% / −31% YTD · 171× / 138× trailing

AMD: alternative-compute optionality capitalized ahead of rack economics. PLTR: excellent workflow franchise, still extraordinary duration. Prefer NVDA and cheaper workflow longs.

Action	Security	Price / YTD	Multiple	Priced-in read	Instruction
Buy	NVDA	$206.84 / +11%	31.7× TTM ~23.8× fwd	Strong growth, meaningful deceleration; little event value from the letter	Core overweight; add below $195–200
Buy	IBM	$214.19 / −27%	19.0×	Low growth and execution skepticism	Value workflow long; build $205–220
Starter	NOW	$98.78 / −36%	61.7×	Growth and AI monetization skepticism, but still expensive	One-third now; add below $90 or on RPO/ACV proof
Spec buy	ORCL	$114.99 / −40%	18.7×	Capex, financing and FCF stress	Half size; add only as backlog converts and funding stabilizes
Accumulate	BOX	$29.01 / −3%	44.4×	Modest growth and uncertain AI monetization	Small position at or below $29
Hold	TSM	$403.41 / +33%	29.8×	Foundry and packaging scarcity	Hold; add after 15%–20% drawdown
Hold / trim	ANET	$173.99 / +33%	59.6×	AI networking growth and share gains	Reduced core; do not chase near high
Trim	VRT	$290.36 / +79%	72.9×	Power/cooling scarcity and premium growth	Take profits; re-enter after 20%+ reset or estimate catch-up
Trim	MU	$920.95 / +223%	20.9×	Severe DRAM/HBM scarcity and earnings upcycle	Tracking position only; low P/E may mark peak earnings
Sell / UW	AMD	$521.95 / +144%	171.3×	Large share gains before full rack-scale proof	Prefer NVDA; relative short with explicit share/economics stop
Sell / UW	PLTR	$122.92 / −31%	138.5×	Exceptional duration still embedded after drawdown	Hedge cheaper workflow longs
Trim / UW	CRWD	$183.28 / +56%	N/M	Platform growth and leverage; not a clean model-COGS trade	Prefer NOW / IBM / BOX on relative price
Avoid	CRWV · NBIS	$71.88 / $187.77 +0% / +124%	N/M / 68.7×	Capacity growth versus financing and residual-value risk	Short basket only when rental, utilization and credit weaken together

Pair 01 / hardware

Long NVDA / short AMD

Delivered systems economics and estimate support versus heavily capitalized catch-up. Cover AMD if rack share and useful-token economics close the gap.

Pair 02 / workflow

Long IBM + NOW + BOX / short PLTR + CRWD

Cheaper control planes and COGS leverage versus premium-duration software. Not factor-neutral; size the short leg smaller.

Pair 03 / infrastructure

Long ORCL / short CRWV + NBIS

Half-gross only: diversified cash flow and contracted demand versus leveraged lessors. Activate after rental, utilization and credit confirmation.

07

CRWV, PLTR, ORCL: demand versus equity capture

The Jensen paper is directionally positive for open-weight deployment. The stock outcome depends on what survives competition, depreciation, financing and the valuation already paid.

The transmission test

More AI demand does not mean more shareholder value.

All three companies can gain workloads in an open-weight world. CoreWeave owns capital-intensive capacity, Palantir owns the workflow layer, and Oracle owns a diversified enterprise and infrastructure stack. Their shareholders receive very different residual economics.

CRWV

Avoid / trigger short

Demand-positive. Capture-fragile.

Open models can lift GPU-hours, but portability makes capacity more comparable and customers more price-sensitive. Equity sits behind debt, leases, depreciation and technology obsolescence.

Q1 revenue$2.08B

Q1 net loss−$740M

Principal debt$25.15B

Top 2 customers65% rev.

Q1 D&A of $1.15B and interest of $536M absorb attractive adjusted EBITDA.
Backlog signals demand, but capacity must be financed and delivered before cash arrives.
Short only when rental prices, utilization/backlog conversion and credit weaken together.

PLTR

Underweight

Business-positive. Valuation-negative.

Model proliferation strengthens the need for ontology, permissions, routing and workflow control. Palantir signed because commoditized models move value toward its layer—not because open weights hurt its franchise.

Q1 growth+85%

GAAP op. margin46%

TTM P/E138.5×

FY26E rev.~37×

The platform is model-neutral across internal, uploaded and hosted models.
The Jensen signal validates the architecture, but years of exceptional execution remain priced.
Underweight the stock, not the operating franchise; cover on material de-rating or estimate outgrowth.

ORCL

Spec buy / half size

Demand-positive. Price-dislocated.

OCI captures inference while databases and applications monetize the surrounding enterprise workload. Sovereign, private and model-neutral deployment fit the paper—and the stock does not carry a PLTR-like premium.

OCI growth+93%

RPO$638B

TTM P/E18.7×

FY27 guide P/E~14.3×

$75B of large AI contracts use customer-prepaid or customer-supplied GPUs.
Diversified operating cash flow cushions the infrastructure build better than a pure lessor.
Half size: capex, negative FCF, debt and backlog conversion remain the real invalidation risks.

Jensen paper / marginal impact

Probability update, not a new thesis.

These scenarios existed before the letter. NVIDIA and the signatories add coordination evidence that customers will choose among models, avoid single-provider lock-in and deploy wherever required. That strengthens routing and inference demand—but also cloud and chip competition.

CRWVMore volume, more portability and price competition. Both sides strengthen; equity capture gets slightly less automatic.

PLTRIts signature validates model-neutral workflow control. The business thesis strengthens; the starting multiple does not improve.

ORCLModel-neutral, sovereign and private inference become more probable. The demand thesis strengthens while financing risk remains unchanged.

Net: long ORCL for dislocated price plus diversified capture; underweight PLTR because the business win is already capitalized; avoid CRWV until rental, utilization and credit prove that token growth reaches equity.

08

Scenario map

The variables that matter are actual weight availability, legal text, hardware capture and premium-spend migration.

Base

Two-tier coexistence

Domestic openness persists; enforcement remains actor-specific. Open volume rises while frontier models retain premium workloads.

Best: workflow, memory, networking
Mixed: NVDA, hyperscalers

Open upside

Kimi validates at scale

Weights ship, independent benchmarks hold, and inference partners make a 3T-class model practical without broad U.S. restriction.

Best: memory, routers, alt compute
Risk: mid-tier closed pricing

Protectionist

Targeted restrictions

Moonshot or related actors face Entity List, diversion or sanctions action. Domestic open weights remain supported.

Best: U.S. labs at margin
Risk: Chinese distribution + compute

Capture downside

Tokens grow; hardware lags

Quantization, batching, old fleets and ASICs reduce cost faster than usage expands. Rental economics and new-GPU intensity disappoint.

Best: application users
Risk: neoclouds, accelerator margins

09

The thesis-changing dashboard

These are the data points that should move the position. Press-release volume and model-download counts should not.

KIMI

27 July catalyst

Weights, license, model card, reproducible serving cost

Change: independent self-host benchmarks on the recommended 64+ accelerator topology.

SPEND

Model economics

Open-weight share of premium workload spend

Change: open models exceed 10% of spend or gain material coding/back-office share.

B200

Rental curve

Methodology-adjusted price versus utilization

Change: decline greater than 15% over 30 days with weak utilization—not a provider-set revision.

HBM

Memory cycle

Contract price, lead time, inventory, capacity

Change: lead-time normalization or supply additions outpacing server/accelerator demand.

NVDA

Q2 FY27

ACIE, networking, margin and power-ready deployment

Change: ACIE/networking slowdown, gross-margin deterioration or customer-financing stress.

LAW

Policy evidence

Actual BIS, OFAC, Commerce or Treasury text

Change: named firms, model classes or distribution channels—not television rhetoric.

Portfolio posture

Buy NVDA + de-rated workflow. Fund it from crowded scarcity.

At the current tape, the preferred expression is long NVIDIA plus IBM, a starter position in ServiceNow, and smaller ORCL/BOX positions. Fund it by trimming MU and VRT, and use AMD and PLTR as the cleanest relative underweights. Neoclouds remain avoids until rental, utilization and credit data jointly confirm the short.

01
Buy: NVDA core; IBM value; NOW starter; ORCL/BOX smaller.
02
Trim: MU and VRT—the bottlenecks are real, but the discovery trade is priced.
03
Underweight: AMD and PLTR, where expectations outrun delivered economics or valuation support.
04
Wait: CRWV/NBIS shorts require falling rental, weak utilization and wider credit spreads.

10

Evidence ledger

Primary sources lead. Secondary reporting is used for chronology, market context and claims not available in issuer material.

01 / Primary

Open weights: the actual trade.

The four-part underwriting frame

Two tiers, not one winner

Models as complements

Tokens ≠ new GPUs

Open at home, targeted abroad

Volume is commoditizing faster than value

The spend gap is the story.

Routing is the equilibrium.

Thin wrappers, not SaaS broadly.

Watch spend, not downloads.

Adopt, modify, reject

The NVIDIA capture equation

One government, two policy vectors

Open, diverse, deployable

Target IP and compute diversion

Where value is most likely to settle

Workflow owners

HBM + server memory

Networking + power

NVIDIA

Neoclouds + wrappers

The priced-in book

Buy delivered economics. Sell capitalized hope.

NVDA

IBM · NOW

MU · VRT

AMD · PLTR

Long NVDA / short AMD

Long IBM + NOW + BOX / short PLTR + CRWD

Long ORCL / short CRWV + NBIS

CRWV, PLTR, ORCL: demand versus equity capture

More AI demand does not mean more shareholder value.

CRWV

PLTR

ORCL

Probability update, not a new thesis.

Scenario map

Two-tier coexistence

Kimi validates at scale

Targeted restrictions

Tokens grow; hardware lags

The thesis-changing dashboard

Weights, license, model card, reproducible serving cost

Open-weight share of premium workload spend

Methodology-adjusted price versus utilization

Contract price, lead time, inventory, capacity

ACIE, networking, margin and power-ready deployment

Actual BIS, OFAC, Commerce or Treasury text

Buy NVDA + de-rated workflow. Fund it from crowded scarcity.

Evidence ledger

NVIDIA open-weights letter

Jensen Huang’s X post

Vercel production index

Kimi K3 launch

Silicon Data GPU indices

Silicon Data announcements

TrendForce Q1 memory

NVIDIA Q1 FY2027

White House NSPM-11

Bessent sanctions comments

NVIDIA tape + valuation

OpenAI Georgia campus

Cross-security valuation tape

CoreWeave Q1 2026

CoreWeave Q1 10-Q

Palantir Q1 2026

Palantir model integration

Oracle Q4 and FY2026